Skip to main content
To KTH's start page To KTH's start page

Biomechanical Parameter Estimation for Wearable Exoskeleton System Design

Time: Fri 2021-10-22 09.00

Location: Sal F3 and via Zoom:, Lindstedtsvägen 26, KTH, Stockholm

Subject area: Engineering Mechanics

Doctoral student: Longbin Zhang , Biomekanik

Opponent: Professor Max Ortiz Catalan,

Supervisor: Professor Elena Gutierrez Farewik, ; Assistant Professor Ruoli Wang, ; Associate professor Christian Smith,

Export to calendar


Exoskeletons are increasingly used in rehabilitation and daily life in persons with motor disorders after neurological injuries. The overall objective of this thesis is to study how to robustly and accurately predict joint torque using inputs from sensors that would technically be feasible to equip on an assistive exoskeleton, and to develop a framework that could be used to evaluate the user-exoskeleton interface. This compilation thesis is based on five papers that focus on different aspects of exoskeleton controller design and physical design, to contribute to the design of wearable exoskeleton systems for rehabilitation and movement assistance.

In the first study, we estimated ankle joint torques using an electromyography (EMG)-driven neuromusculoskeletal (NMS) model and a standard artificial neural network (ANN) model in seven movement tasks. EMG signals and ankle joint angles were used as input in each method, the models were trained with experimental data from 3D motion analysis, and ankle joint torques were predicted. Our results suggested that the standard-ANN model could predict ankle torques more accurately when trained on a large and varied set of trials but less accurately in unseen movements whereas the NMS model was overall more useful and has the advantage of detailing underlying physical principles.

The second study was directly inspired by the first study. Our aim was to estimate ankle joint torques using a hybrid form of the two methods in the first study, specifically an NMS solver-informed ANN, by augmenting a standard-ANN model with additional features from the underlying NMS solver. The torque prediction performance of hybrid-ANN models was investigated in two scenarios: intra-subject and inter-subject tests. We found that the NMS solver informed-ANN model had a better joint torque prediction accuracy than the methods independently (i.e. NMS or the standard-ANN model), which indicates that benefit was gained from integrating NMS features into standard ANN models. Furthermore, in the inter-subject tests, we adopted the transfer learning technique to further improve the joint torque performance by taking advantage of the information extracted from previous subjects.  

In the third study, we predicted not only ankle joint torques but also knee joint torques in the sagittal plane and hip joint torques in both sagittal and frontal planes during several different motions using long short-term memory (LSTM) neural networks. Several networks were constructed, namely separate LSTM models for each movement as well as one uniform model for all movements. The LSTM networks were constructed both with and without transfer learning, in which the networks could take advantage of information extracted from previous tasks and/or subjects to predict for an unseen or ``new'' task or subject. We evaluated the LSTM models' prediction ability across tasks and subjects and studied whether generalizability was improved when transfer learning was implemented in the LSTM networks.  Our results indicate that both one uniform and ten separate LSTM models predicted lower limb joint torques accurately, i.e. with low prediction error, in intra-subject tests. By including transfer learning, the LSTM models' generalizability in predicting moments across tasks and subjects was significantly improved, even if re-trained with a smaller subset of movements from the ``new'' subject. 

In the fourth and fifth studies, we generated a realistic knee exoskeleton model. In the fourth study, we created a virtual human-machine system (HMS) in a musculoskeletal modeling software. We found that this proposed virtual HMS was useful for simulating how different assistive strategies may affect parameters that describe the demands and consequences on the user, including muscular demand, joint contact forces and human-machine interactive forces involved in movements. In the fifth study, we analyzed, through simulation, how different weight distributions of knee exoskeleton components influenced the user's muscle activation during three functional movements. One unilateral knee exoskeleton prototype was fabricated and tested on five healthy subjects. Simulation and experimental muscle activations were compared. We found that muscle activation varied among weight distributions and movements, indicating that no single physical design was optimal for all movements. Exoskeleton physical design should ideally consider the user's activity goals. 

These methods developed in this thesis hold great promise in applications such as the design of assistance-as-needed exoskeleton control. The developed virtual HMS makes it possible to study how a knee exoskeleton's different assistive strategies and physical designs can influence its user's muscle effort and interaction forces, which can both simplify prototype iterations and evaluate some parameters that are otherwise difficult to measure experimentally.